Prediction errors disrupt hippocampal representations and update episodic memories

The brain supports adaptive behavior by generating predictions, learning from errors, and updating memories to incorporate new information. Prediction error, or surprise, triggers learning when reality contradicts expectations. Prior studies have shown that the hippocampus signals prediction errors, but the hypothesized link to memory updating has not been demonstrated. In a human functional MRI study, we elicited mnemonic prediction errors by interrupting familiar narrative videos immediately before the expected endings. We found that prediction errors reversed the relationship between univariate hippocampal activation and memory: greater hippocampal activation predicted memory preservation after expected endings, but memory updating after surprising endings. In contrast to previous studies, we show that univariate activation was insufficient for understanding hippocampal prediction error signals. We explain this surprising finding by tracking both the evolution of hippocampal activation patterns and the connectivity between the hippocampus and neuromodulatory regions. We found that hippocampal activation patterns stabilized as each narrative episode unfolded, suggesting sustained episodic representations. Prediction errors disrupted these sustained representations and the degree of disruption predicted memory updating. The relationship between hippocampal activation and subsequent memory depended on concurrent basal forebrain activation, supporting the idea that cholinergic modulation regulates attention and memory. We conclude that prediction errors create conditions that favor memory updating, prompting the hippocampus to abandon ongoing predictions and make memories malleable.

In daily life, we continuously draw on past experiences to predict the future. Expectation and surprise shape learning across many situations, such as when we discover misinformation in the news, receive feedback on an examination, or make decisions based on past outcomes. When our predictions are incorrect, we must update our mnemonic models of the world to support adaptive behavior. Prediction error is a measure of the discrepancy between expectation and reality; this surprise signal is both evident in brain activity and related to learning (1–6). The brain dynamically reconstructs memories during recall, recreating and revising past experiences based on current information (7). The intuitive idea that surprise governs learning has long shaped our understanding of memory, reward learning, perception, action, and social behavior (2, 8–14). Yet, the neural mechanisms that allow prediction error to update memories remain unknown.Past research has implicated the hippocampus in each of the mnemonic functions required for learning from prediction errors: retrieving memories to make predictions, identifying discrepancies between past and present, and encoding new information (2, 15–20). Functional MRI (fMRI) studies have shown that hippocampal activation increases after predictions are violated; this surprise response has been termed “mismatch detection” (18, 19, 21–23) or “mnemonic prediction error” (20). These past studies have shown that the hippocampus detects mnemonic prediction errors. Several theoretical frameworks have hypothesized that this hippocampal prediction error signal could update memories (17, 20, 24–27), but this crucial link for understanding how we learn from error has not yet been demonstrated.What mechanisms could link hippocampal prediction errors to memory updating? A leading hypothesis is that prediction errors shift the focus of attention and adjust cognitive processing (20, 28–32). After episodes that align with expectations, we should continue generating predictions and shift attention internally, sustaining and reinforcing existing memories. However, after mnemonic prediction errors, we should reset our expectations and shift attention externally, preparing to encode new information and update memories. Consistent with this idea, mnemonic prediction errors have been shown to enhance the hippocampal input pathway that supports encoding, but suppress the output pathway that supports retrieval (20). We propose that surprising events may also change intrinsic hippocampal processing, changing the effect of hippocampal activation on memory outcomes.Neuromodulation may be a critical factor that regulates hippocampal processing and enables memory updating. Currently, there is mixed evidence supporting two hypotheses: acetylcholine or dopamine could act upon the hippocampus to regulate processing after surprising events (24–27, 29, 31, 33, 34). Several models have proposed that acetylcholine from the medial septum (within the basal forebrain) regulates the balance between input and output pathways in the hippocampus (27–29, 35–38), thus allowing stored memories to be compared with perceptual input (31, 38, 39). After prediction errors, acetylcholine release could change hippocampal processing and enhance encoding or memory updating (26, 29, 33, 37, 39). On the other hand, dopamine released from the ventral tegmental area (VTA), if transmitted to the hippocampus, could also modulate hippocampal plasticity after prediction errors. Past studies have shown that the hippocampus and VTA are coactivated after surprising events (40, 41). Other work has shown that coactivation of the hippocampus and VTA predicts memory encoding and integration (42–45). Overall, basal forebrain and VTA neuromodulation are both candidate mechanisms for regulating hippocampal processing and memory updating.In the present study, we used an fMRI task with human participants to examine trial-wise hippocampal responses to prediction errors during narrative videos. During the “encoding phase,” participants viewed 70 full-length videos that featured narrative episodes with salient endings (e.g., a baseball batter hitting a home run) (Fig. 1A). During the “reactivation phase” the following day, participants watched the videos again (Fig. 1B). We elicited mnemonic prediction errors by interrupting half of the videos immediately before the expected narrative ending (e.g., the video ends while the baseball batter is midswing). These surprising interruptions were comparable to the prediction errors employed in prior studies of memory updating (1). Half of the videos were presented in full-length form (Full, as previously seen during the encoding phase) and half were presented in interrupted form (Interrupted, eliciting prediction error).Open in a separate windowFig. 1.Overview of experimental paradigm. (A) During the encoding phase, all videos were presented in full-length form. Here we show example frames depicting a stimulus video. (B) During the reactivation phase, participants viewed the 70 videos again, but half (35 videos) were interrupted to elicit mnemonic prediction error. Participants were cued with the video name, watched the video (Full or Interrupted), and then viewed a fixation screen. The “baseball” video was interrupted when the batter was midswing. fMRI analyses focused on the postvideo fixation periods (red highlighted boxes). Thus, visual and auditory stimulation were matched across Full and Interrupted conditions, allowing us to compare postvideo neural activation while controlling for perceptual input. (C) During the test phase, participants answered structured interview questions about all 70 videos, and were instructed to answer based on their memory of the Full video originally shown during the Encoding phase. Here we show example text illustrating the memory test format and scoring of correct details (our measure of memory preservation) and false memories (our measure of memory updating, because false memories indicate that the memory has been modified). The void response (“I don’t remember”) is not counted as a false memory. (D) Overview of the experiment. All participants completed encoding, reactivation, and test phases of the study. The Delayed group (fMRI participants) completed the test phase 24 h after reactivation, because prior studies have shown that memory updating becomes evident only after a delay (e.g., to permit protein synthesis). The Immediate group completed the test phase immediately after reactivation and was not scanned. The purpose of the Immediate group was to test the behavioral prediction that memory updating required a delay.During the “test phase,” participants completed a memory test in the form of a structured interview (Fig. 1C). On each trial, participants were cued with the name of the video and recalled the narrative. The experimenter then probed for further details with predetermined questions (e.g., “Can you describe the baseball batter’s ethnicity, age range, or clothing?”). Our critical measure of memory updating was “false memories,” because the presence of a false memory indicates that the original memory was changed in some way. Although it can be adaptive to update real-world memories by incorporating relevant new information, we expected that our laboratory paradigm would induce false memories because participants would integrate interfering details across similar episodes (1, 7). Because we were interested in false memories as a measure of memory updating, we instructed participants not to guess and permitted them to skip details they could not recall.Prior research in human and animals has shown that some memory-updating effects only emerge after delays that allow protein synthesis to occur during consolidation and reconsolidation (1, 46–48). Therefore, to test our primary question about the neural correlates of memory updating, fMRI participants completed the encoding, reactivation, and test phases over 3 d, with 24-h between each session (Delayed group, n = 24). In addition, we tested the behavioral prediction that memory updating would require a delay (i.e., because transforming a memory trace requires protein synthesis) by recruiting a separate group of participants who completed the test phase immediately after the reactivation phase on day 2 (Immediate group, n = 24) (Fig. 1D). Delayed group participants completed the reactivation phase while undergoing an fMRI scan, whereas Immediate group participants (n = 24) were not scanned. Our primary fMRI analyses examined the fixation period immediately following the offset of Full and Interrupted videos (postvideo period) (Fig. 1 B, Right) during the reactivation phase in the Delayed group. Importantly, this design compares neural responses to surprising and expected video endings while controlling for visual and auditory input.Our approach allowed us to test several questions set up by the prior literature. First, we used naturalistic video stimuli to examine the effect of mnemonic prediction error on hippocampal activation and episodic memories. Second, to investigate hippocampal processing, we used multivariate analyses to track how episodic representations were sustained or disrupted over time. Third, to test hypotheses about neuromodulatory mechanisms, we related hippocampal activation and memory updating to activation in the basal forebrain and VTA.